Protocol 08: Target Trial for Chatbot-Linked Crisis
Question
Does a sharp increase in generative-chatbot exposure raise near-term risk of psychotic or manic symptoms beyond sleep loss, substance use, acute stress, and pre-existing vulnerability?
Why this protocol exists
Case reports can establish chronology and reveal plausible amplification. They cannot, by themselves, isolate chatbot exposure from the conditions that caused both intense use and crisis. The missing object is a repeated, time-ordered comparison with explicit rival causes.
Emulated trial
Eligibility
Participants with regular chatbot access, no acute psychotic episode at baseline, and capacity to complete repeated assessments. Record prior psychosis or mania, family history, medication changes, baseline symptom burden, and ordinary chatbot-use level before follow-up.
Time zero
Each participant-day begins at a fixed local time. Exposure, covariates, and outcomes must be assigned to non-overlapping windows so that an outcome cannot be used to define its own supposed cause.
Exposure strategies
Compare person-days with:
- Exposure spike: chatbot minutes or messages substantially above that person's established baseline.
- No spike: use within that person's ordinary range.
- Feature-specific spike: unusually high nocturnal use, repeated reassurance seeking, or a high proportion of affirming responses.
Thresholds must be fixed before outcome analysis. Testing many thresholds and publishing only the dramatic one manufactures a pattern.
Outcomes
Primary outcome: independently assessed increase in validated psychosis- or mania-related symptoms during a predeclared lag window.
Secondary outcomes:
- urgent psychiatric contact;
- functional impairment;
- sleep reduction;
- conviction in a pre-specified unusual belief;
- escalation in reassurance-seeking dialogue.
Chat transcripts should not be shown to outcome raters during the first assessment pass.
Minimum repeated measurements
Time-stamp, at least daily:
- chatbot minutes and messages;
- nocturnal use;
- model and product changes;
- user prompts and response features coded by blinded raters;
- sleep duration and timing;
- alcohol and other substance exposure;
- medication adherence and changes;
- acute stressors;
- symptoms and functioning.
Passive traces may improve timing, but they do not convert a correlation into a cause.
Analysis
Use within-person estimates so stable vulnerability is not mistaken for an exposure effect. Adjust for time-varying sleep, substances, stress, medication changes, and prior symptoms. Test several predeclared lag structures, including whether symptoms rise before exposure spikes.
A negative-control exposure could be intensive use of a non-conversational digital tool at the same hour. A negative-control outcome should be chosen that chatbot agreement should not plausibly affect but sleep loss might.
Rival predictions
| Explanation | Expected sequence | |---|---| | Chatbot-specific ignition | Exposure or affirming-response spike precedes symptoms after rival causes are controlled | | Amplification | Symptoms or unusual beliefs begin first; affirming interaction predicts subsequent conviction or persistence | | Reverse causation | Symptoms reliably predict later use spikes more strongly than use predicts later symptoms | | Common cause | Sleep loss, substances, stress, or mania predict both use and symptoms; chatbot coefficient attenuates sharply after adjustment | | Reporting selection | Strong effects appear mainly in cases selected because transcripts are dramatic, not in prospective sampling |
Evidence that would change the ledger
The chatbot-specific causal claim gains weight if exposure spikes repeatedly precede symptom increases within the same people, show a dose-response or feature-response pattern, survive adjustment for measured rivals, and weaken when the exposure stops while rivals remain stable.
It loses weight if symptoms lead exposure, if adjustment for sleep or substances collapses the association, if non-conversational nocturnal use produces the same signal, or if prospective cohorts show no discriminating exposure pattern.
Current status
This is a discriminating design, not evidence that the effect exists. Existing retrieved material supports feasibility of dense longitudinal measurement and plausibility of AI agreement and overreliance mechanisms. It does not yet supply the prospective causal result this protocol requires.
Open questions
- What validated symptom measure is sensitive enough for daily use without turning ordinary eccentricity into pathology?
- Can response affirmation be coded reliably without leaking symptom severity to raters?
- What lag window is clinically plausible and resistant to researcher degrees of freedom?
- How should emergency escalation be handled without contaminating observation or withholding care?